| The purpose of person re-identification(Re ID)is to recognize related person images under multiple cameras based on a person query image.Although a static model of supervised learning training performs well on specified annotated data sets,but such methods can’t be adequately adapted to the complex scene of constant construction and rely on manual annotation of person data.Due to the above limitations,research has gradually turned to domain adaptive Re ID.Based on this direction,this paper firstly elaborates,combs and analyzes the research on Re ID at home and abroad in recent years,and finds that this field still has important research significance.Among them,the problems of person occlusion,low resolution(Low-Res),style differences between cameras and domains,global and local feature differences,and low fault tolerance rate of model training still face great challenges.The main research contents and results are as follows:(1)This paper proposes a domain adaptive Re ID method DMS based on feature enhancement and style diversity,which includes three main contents: Firstly,aiming at the problems of person occlusion and Low-Res in the research field,network structure design based on Deformable Convolution and Attention Mechanisms(D-AM)is proposed.In the process of feature extraction,this method realizes the adaptive sampling of network model and the enhancement of key information about person,and completes the feature enhancement process.Then,Sample Style Augmentation(SSA)model is given to reflect the style differences among different cameras.SSA realizes sample style enhancement based on generative adversarial network,proposes the person difference loss between cameras to dig the differences between cameras deeply,and promotes the style diversity of network learning SSA model.At the same time,a Multi-loss Training(ML)method is proposed to coordinate the training optimization of the model,excavate the deep characteristics of person,and further improve the performance of the model.The experimental results show that D-AM,ML and SSA can improve the performance of the model.When used in combination,the performance of the model is superior,which surpasses some of the current mainstream methods,and verifies the effectiveness and advancement of the method.(2)On the basis of DMS,a domain adaptive Re ID method DMS-SC based on composite label and feature division is proposed.The improvement of this method includes three main contents: Firstly,to solve the problems of domain differences and insufficient source domain data mining,Style Transfer Hybrid Memory(STHM)is proposed.This method fully mines the characteristics of source domain data after style migration,dynamically updates the prototype of target domain and source domain category,and improves the model performance.Then,aiming at the difference between global and local features,a multi-dimensional local feature partitioning rule is proposed to promote the model to pay attention to the difference and connection between global and local features,and reflect the specificity of person features under different classification rules.At the same time,collaborative optimization strategy guided by Composite Label(CL)is proposed to solve the problem of low fault tolerance in model training caused by false label noise.This strategy implements the application of multi-dimension local feature partitioning rules in a multi-branch network,alleviates the problem of difference caused by focusing only on global features,and provides more reliable guidance for model training.Experimental results show that STHM,CL can improve model performance,DMS has good universality,verify the effectiveness of multi-dimensional local feature division rules and division network design,proved that DMS-SC method has the effectiveness,advancement extensibility and excellent performance,beyond the current part of the mainstream methods. |
